Industry coalition introduces new benchmark to rate safety of AI models

The artificial intelligence industry has reached a significant milestone with the introduction of a standardized benchmark system designed to evaluate the potential risks and harmful behaviors of AI language models.

New industry standard: MLCommons, a nonprofit organization with 125 member organizations including major tech companies and academic institutions, has launched AILuminate, a comprehensive benchmark system for assessing AI safety risks.

• The benchmark tests AI models against more than 12,000 prompts across 12 categories, including violent crime incitement, child exploitation, hate speech, and intellectual property infringement
• Models receive ratings ranging from “poor” to “excellent” based on their performance
• Test prompts remain confidential to prevent AI models from being trained specifically to pass the evaluations

Initial testing results: Several prominent AI companies have already subjected their models to AILuminate’s evaluation process, revealing varying levels of safety performance.

• Anthropic’s Claude, Google’s Gemma, and Microsoft’s Phi achieved “very good” ratings
• OpenAI’s GPT-4 and Meta’s largest Llama model received “good” ratings
• The Allen Institute for AI’s OLMo scored “poor,” though it was noted as being a research model not designed with safety features

Global implications: The benchmark system represents a step toward international standardization of AI safety measurements and accountability.

• Chinese companies Huawei and Alibaba are among MLCommons’ member organizations, though none have yet used the new benchmark
• MLCommons has partnered with Singapore-based AI Verify to develop standards incorporating Asian perspectives
• The system could provide a way to compare AI safety standards across different countries and regions

Political context: The timing of this benchmark’s introduction coincides with uncertainty around future AI regulation in the United States.

• Donald Trump has promised to eliminate President Biden’s AI Executive Order if elected
• The current executive order established an AI Safety Institute and introduced corporate responsibility measures
• MLCommons aims to maintain industry standards regardless of political changes

Looking forward: While AILuminate represents a significant advance in AI safety evaluation, it addresses only certain aspects of AI risk.

• The benchmark does not measure potential risks related to AI deception or control issues
• MLCommons plans to evolve the standards over time, similar to automotive safety ratings
• The organization’s agility may allow it to adapt more quickly to emerging AI developments than government regulators

Industry perspective and implications: This new benchmark system could reshape how AI companies approach safety testing and development.

• The standardized evaluation process may encourage companies to prioritize safety features in their AI models
• Results could influence market competition and consumer trust in AI products
• The system’s success will depend on widespread adoption and continued evolution to address emerging AI challenges